1 // SPDX-License-Identifier: GPL-2.0 2 #include <linux/kernel.h> 3 #include <linux/errno.h> 4 #include <linux/fs.h> 5 #include <linux/file.h> 6 #include <linux/mm.h> 7 #include <linux/slab.h> 8 #include <linux/nospec.h> 9 #include <linux/hugetlb.h> 10 #include <linux/compat.h> 11 #include <linux/io_uring.h> 12 13 #include <uapi/linux/io_uring.h> 14 15 #include "io_uring.h" 16 #include "openclose.h" 17 #include "rsrc.h" 18 19 struct io_rsrc_update { 20 struct file *file; 21 u64 arg; 22 u32 nr_args; 23 u32 offset; 24 }; 25 26 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, 27 struct io_mapped_ubuf **pimu, 28 struct page **last_hpage); 29 30 /* only define max */ 31 #define IORING_MAX_FIXED_FILES (1U << 20) 32 #define IORING_MAX_REG_BUFFERS (1U << 14) 33 34 static inline bool io_put_rsrc_data_ref(struct io_rsrc_data *rsrc_data) 35 { 36 return !--rsrc_data->refs; 37 } 38 39 int __io_account_mem(struct user_struct *user, unsigned long nr_pages) 40 { 41 unsigned long page_limit, cur_pages, new_pages; 42 43 if (!nr_pages) 44 return 0; 45 46 /* Don't allow more pages than we can safely lock */ 47 page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT; 48 49 cur_pages = atomic_long_read(&user->locked_vm); 50 do { 51 new_pages = cur_pages + nr_pages; 52 if (new_pages > page_limit) 53 return -ENOMEM; 54 } while (!atomic_long_try_cmpxchg(&user->locked_vm, 55 &cur_pages, new_pages)); 56 return 0; 57 } 58 59 static void io_unaccount_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) 60 { 61 if (ctx->user) 62 __io_unaccount_mem(ctx->user, nr_pages); 63 64 if (ctx->mm_account) 65 atomic64_sub(nr_pages, &ctx->mm_account->pinned_vm); 66 } 67 68 static int io_account_mem(struct io_ring_ctx *ctx, unsigned long nr_pages) 69 { 70 int ret; 71 72 if (ctx->user) { 73 ret = __io_account_mem(ctx->user, nr_pages); 74 if (ret) 75 return ret; 76 } 77 78 if (ctx->mm_account) 79 atomic64_add(nr_pages, &ctx->mm_account->pinned_vm); 80 81 return 0; 82 } 83 84 static int io_copy_iov(struct io_ring_ctx *ctx, struct iovec *dst, 85 void __user *arg, unsigned index) 86 { 87 struct iovec __user *src; 88 89 #ifdef CONFIG_COMPAT 90 if (ctx->compat) { 91 struct compat_iovec __user *ciovs; 92 struct compat_iovec ciov; 93 94 ciovs = (struct compat_iovec __user *) arg; 95 if (copy_from_user(&ciov, &ciovs[index], sizeof(ciov))) 96 return -EFAULT; 97 98 dst->iov_base = u64_to_user_ptr((u64)ciov.iov_base); 99 dst->iov_len = ciov.iov_len; 100 return 0; 101 } 102 #endif 103 src = (struct iovec __user *) arg; 104 if (copy_from_user(dst, &src[index], sizeof(*dst))) 105 return -EFAULT; 106 return 0; 107 } 108 109 static int io_buffer_validate(struct iovec *iov) 110 { 111 unsigned long tmp, acct_len = iov->iov_len + (PAGE_SIZE - 1); 112 113 /* 114 * Don't impose further limits on the size and buffer 115 * constraints here, we'll -EINVAL later when IO is 116 * submitted if they are wrong. 117 */ 118 if (!iov->iov_base) 119 return iov->iov_len ? -EFAULT : 0; 120 if (!iov->iov_len) 121 return -EFAULT; 122 123 /* arbitrary limit, but we need something */ 124 if (iov->iov_len > SZ_1G) 125 return -EFAULT; 126 127 if (check_add_overflow((unsigned long)iov->iov_base, acct_len, &tmp)) 128 return -EOVERFLOW; 129 130 return 0; 131 } 132 133 static void io_buffer_unmap(struct io_ring_ctx *ctx, struct io_mapped_ubuf **slot) 134 { 135 struct io_mapped_ubuf *imu = *slot; 136 unsigned int i; 137 138 if (imu != ctx->dummy_ubuf) { 139 for (i = 0; i < imu->nr_bvecs; i++) 140 unpin_user_page(imu->bvec[i].bv_page); 141 if (imu->acct_pages) 142 io_unaccount_mem(ctx, imu->acct_pages); 143 kvfree(imu); 144 } 145 *slot = NULL; 146 } 147 148 static void io_rsrc_put_work_one(struct io_rsrc_data *rsrc_data, 149 struct io_rsrc_put *prsrc) 150 { 151 struct io_ring_ctx *ctx = rsrc_data->ctx; 152 153 if (prsrc->tag) 154 io_post_aux_cqe(ctx, prsrc->tag, 0, 0); 155 rsrc_data->do_put(ctx, prsrc); 156 } 157 158 static void __io_rsrc_put_work(struct io_rsrc_node *ref_node) 159 { 160 struct io_rsrc_data *rsrc_data = ref_node->rsrc_data; 161 struct io_rsrc_put *prsrc, *tmp; 162 163 if (ref_node->inline_items) 164 io_rsrc_put_work_one(rsrc_data, &ref_node->item); 165 166 list_for_each_entry_safe(prsrc, tmp, &ref_node->item_list, list) { 167 list_del(&prsrc->list); 168 io_rsrc_put_work_one(rsrc_data, prsrc); 169 kfree(prsrc); 170 } 171 172 io_rsrc_node_destroy(rsrc_data->ctx, ref_node); 173 if (io_put_rsrc_data_ref(rsrc_data)) 174 complete(&rsrc_data->done); 175 } 176 177 void io_wait_rsrc_data(struct io_rsrc_data *data) 178 { 179 if (data && !io_put_rsrc_data_ref(data)) 180 wait_for_completion(&data->done); 181 } 182 183 void io_rsrc_node_destroy(struct io_ring_ctx *ctx, struct io_rsrc_node *node) 184 { 185 if (!io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache)) 186 kfree(node); 187 } 188 189 void io_rsrc_node_ref_zero(struct io_rsrc_node *node) 190 __must_hold(&node->rsrc_data->ctx->uring_lock) 191 { 192 struct io_ring_ctx *ctx = node->rsrc_data->ctx; 193 194 node->done = true; 195 while (!list_empty(&ctx->rsrc_ref_list)) { 196 node = list_first_entry(&ctx->rsrc_ref_list, 197 struct io_rsrc_node, node); 198 /* recycle ref nodes in order */ 199 if (!node->done) 200 break; 201 202 list_del(&node->node); 203 __io_rsrc_put_work(node); 204 } 205 } 206 207 struct io_rsrc_node *io_rsrc_node_alloc(struct io_ring_ctx *ctx) 208 { 209 struct io_rsrc_node *ref_node; 210 struct io_cache_entry *entry; 211 212 entry = io_alloc_cache_get(&ctx->rsrc_node_cache); 213 if (entry) { 214 ref_node = container_of(entry, struct io_rsrc_node, cache); 215 } else { 216 ref_node = kzalloc(sizeof(*ref_node), GFP_KERNEL); 217 if (!ref_node) 218 return NULL; 219 } 220 221 ref_node->rsrc_data = NULL; 222 ref_node->refs = 1; 223 INIT_LIST_HEAD(&ref_node->node); 224 INIT_LIST_HEAD(&ref_node->item_list); 225 ref_node->done = false; 226 ref_node->inline_items = 0; 227 return ref_node; 228 } 229 230 void io_rsrc_node_switch(struct io_ring_ctx *ctx, 231 struct io_rsrc_data *data_to_kill) 232 __must_hold(&ctx->uring_lock) 233 { 234 struct io_rsrc_node *node = ctx->rsrc_node; 235 struct io_rsrc_node *backup = io_rsrc_node_alloc(ctx); 236 237 if (WARN_ON_ONCE(!backup)) 238 return; 239 240 data_to_kill->refs++; 241 node->rsrc_data = data_to_kill; 242 list_add_tail(&node->node, &ctx->rsrc_ref_list); 243 /* put master ref */ 244 io_put_rsrc_node(ctx, node); 245 ctx->rsrc_node = backup; 246 } 247 248 int io_rsrc_node_switch_start(struct io_ring_ctx *ctx) 249 { 250 if (io_alloc_cache_empty(&ctx->rsrc_node_cache)) { 251 struct io_rsrc_node *node = kzalloc(sizeof(*node), GFP_KERNEL); 252 253 if (!node) 254 return -ENOMEM; 255 io_alloc_cache_put(&ctx->rsrc_node_cache, &node->cache); 256 } 257 return 0; 258 } 259 260 __cold static int io_rsrc_ref_quiesce(struct io_rsrc_data *data, 261 struct io_ring_ctx *ctx) 262 { 263 int ret; 264 265 /* As we may drop ->uring_lock, other task may have started quiesce */ 266 if (data->quiesce) 267 return -ENXIO; 268 ret = io_rsrc_node_switch_start(ctx); 269 if (ret) 270 return ret; 271 io_rsrc_node_switch(ctx, data); 272 273 /* kill initial ref */ 274 if (io_put_rsrc_data_ref(data)) 275 return 0; 276 277 data->quiesce = true; 278 mutex_unlock(&ctx->uring_lock); 279 do { 280 ret = io_run_task_work_sig(ctx); 281 if (ret < 0) { 282 mutex_lock(&ctx->uring_lock); 283 if (!data->refs) { 284 ret = 0; 285 } else { 286 /* restore the master reference */ 287 data->refs++; 288 } 289 break; 290 } 291 ret = wait_for_completion_interruptible(&data->done); 292 if (!ret) { 293 mutex_lock(&ctx->uring_lock); 294 if (!data->refs) 295 break; 296 /* 297 * it has been revived by another thread while 298 * we were unlocked 299 */ 300 mutex_unlock(&ctx->uring_lock); 301 } 302 } while (1); 303 data->quiesce = false; 304 305 return ret; 306 } 307 308 static void io_free_page_table(void **table, size_t size) 309 { 310 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); 311 312 for (i = 0; i < nr_tables; i++) 313 kfree(table[i]); 314 kfree(table); 315 } 316 317 static void io_rsrc_data_free(struct io_rsrc_data *data) 318 { 319 size_t size = data->nr * sizeof(data->tags[0][0]); 320 321 if (data->tags) 322 io_free_page_table((void **)data->tags, size); 323 kfree(data); 324 } 325 326 static __cold void **io_alloc_page_table(size_t size) 327 { 328 unsigned i, nr_tables = DIV_ROUND_UP(size, PAGE_SIZE); 329 size_t init_size = size; 330 void **table; 331 332 table = kcalloc(nr_tables, sizeof(*table), GFP_KERNEL_ACCOUNT); 333 if (!table) 334 return NULL; 335 336 for (i = 0; i < nr_tables; i++) { 337 unsigned int this_size = min_t(size_t, size, PAGE_SIZE); 338 339 table[i] = kzalloc(this_size, GFP_KERNEL_ACCOUNT); 340 if (!table[i]) { 341 io_free_page_table(table, init_size); 342 return NULL; 343 } 344 size -= this_size; 345 } 346 return table; 347 } 348 349 __cold static int io_rsrc_data_alloc(struct io_ring_ctx *ctx, 350 rsrc_put_fn *do_put, u64 __user *utags, 351 unsigned nr, struct io_rsrc_data **pdata) 352 { 353 struct io_rsrc_data *data; 354 int ret = 0; 355 unsigned i; 356 357 data = kzalloc(sizeof(*data), GFP_KERNEL); 358 if (!data) 359 return -ENOMEM; 360 data->tags = (u64 **)io_alloc_page_table(nr * sizeof(data->tags[0][0])); 361 if (!data->tags) { 362 kfree(data); 363 return -ENOMEM; 364 } 365 366 data->nr = nr; 367 data->ctx = ctx; 368 data->do_put = do_put; 369 data->refs = 1; 370 if (utags) { 371 ret = -EFAULT; 372 for (i = 0; i < nr; i++) { 373 u64 *tag_slot = io_get_tag_slot(data, i); 374 375 if (copy_from_user(tag_slot, &utags[i], 376 sizeof(*tag_slot))) 377 goto fail; 378 } 379 } 380 init_completion(&data->done); 381 *pdata = data; 382 return 0; 383 fail: 384 io_rsrc_data_free(data); 385 return ret; 386 } 387 388 static int __io_sqe_files_update(struct io_ring_ctx *ctx, 389 struct io_uring_rsrc_update2 *up, 390 unsigned nr_args) 391 { 392 u64 __user *tags = u64_to_user_ptr(up->tags); 393 __s32 __user *fds = u64_to_user_ptr(up->data); 394 struct io_rsrc_data *data = ctx->file_data; 395 struct io_fixed_file *file_slot; 396 struct file *file; 397 int fd, i, err = 0; 398 unsigned int done; 399 bool needs_switch = false; 400 401 if (!ctx->file_data) 402 return -ENXIO; 403 if (up->offset + nr_args > ctx->nr_user_files) 404 return -EINVAL; 405 406 for (done = 0; done < nr_args; done++) { 407 u64 tag = 0; 408 409 if ((tags && copy_from_user(&tag, &tags[done], sizeof(tag))) || 410 copy_from_user(&fd, &fds[done], sizeof(fd))) { 411 err = -EFAULT; 412 break; 413 } 414 if ((fd == IORING_REGISTER_FILES_SKIP || fd == -1) && tag) { 415 err = -EINVAL; 416 break; 417 } 418 if (fd == IORING_REGISTER_FILES_SKIP) 419 continue; 420 421 i = array_index_nospec(up->offset + done, ctx->nr_user_files); 422 file_slot = io_fixed_file_slot(&ctx->file_table, i); 423 424 if (file_slot->file_ptr) { 425 file = (struct file *)(file_slot->file_ptr & FFS_MASK); 426 err = io_queue_rsrc_removal(data, i, ctx->rsrc_node, file); 427 if (err) 428 break; 429 file_slot->file_ptr = 0; 430 io_file_bitmap_clear(&ctx->file_table, i); 431 needs_switch = true; 432 } 433 if (fd != -1) { 434 file = fget(fd); 435 if (!file) { 436 err = -EBADF; 437 break; 438 } 439 /* 440 * Don't allow io_uring instances to be registered. If 441 * UNIX isn't enabled, then this causes a reference 442 * cycle and this instance can never get freed. If UNIX 443 * is enabled we'll handle it just fine, but there's 444 * still no point in allowing a ring fd as it doesn't 445 * support regular read/write anyway. 446 */ 447 if (io_is_uring_fops(file)) { 448 fput(file); 449 err = -EBADF; 450 break; 451 } 452 err = io_scm_file_account(ctx, file); 453 if (err) { 454 fput(file); 455 break; 456 } 457 *io_get_tag_slot(data, i) = tag; 458 io_fixed_file_set(file_slot, file); 459 io_file_bitmap_set(&ctx->file_table, i); 460 } 461 } 462 463 if (needs_switch) 464 io_rsrc_node_switch(ctx, data); 465 return done ? done : err; 466 } 467 468 static int __io_sqe_buffers_update(struct io_ring_ctx *ctx, 469 struct io_uring_rsrc_update2 *up, 470 unsigned int nr_args) 471 { 472 u64 __user *tags = u64_to_user_ptr(up->tags); 473 struct iovec iov, __user *iovs = u64_to_user_ptr(up->data); 474 struct page *last_hpage = NULL; 475 bool needs_switch = false; 476 __u32 done; 477 int i, err; 478 479 if (!ctx->buf_data) 480 return -ENXIO; 481 if (up->offset + nr_args > ctx->nr_user_bufs) 482 return -EINVAL; 483 484 for (done = 0; done < nr_args; done++) { 485 struct io_mapped_ubuf *imu; 486 int offset = up->offset + done; 487 u64 tag = 0; 488 489 err = io_copy_iov(ctx, &iov, iovs, done); 490 if (err) 491 break; 492 if (tags && copy_from_user(&tag, &tags[done], sizeof(tag))) { 493 err = -EFAULT; 494 break; 495 } 496 err = io_buffer_validate(&iov); 497 if (err) 498 break; 499 if (!iov.iov_base && tag) { 500 err = -EINVAL; 501 break; 502 } 503 err = io_sqe_buffer_register(ctx, &iov, &imu, &last_hpage); 504 if (err) 505 break; 506 507 i = array_index_nospec(offset, ctx->nr_user_bufs); 508 if (ctx->user_bufs[i] != ctx->dummy_ubuf) { 509 err = io_queue_rsrc_removal(ctx->buf_data, i, 510 ctx->rsrc_node, ctx->user_bufs[i]); 511 if (unlikely(err)) { 512 io_buffer_unmap(ctx, &imu); 513 break; 514 } 515 ctx->user_bufs[i] = ctx->dummy_ubuf; 516 needs_switch = true; 517 } 518 519 ctx->user_bufs[i] = imu; 520 *io_get_tag_slot(ctx->buf_data, i) = tag; 521 } 522 523 if (needs_switch) 524 io_rsrc_node_switch(ctx, ctx->buf_data); 525 return done ? done : err; 526 } 527 528 static int __io_register_rsrc_update(struct io_ring_ctx *ctx, unsigned type, 529 struct io_uring_rsrc_update2 *up, 530 unsigned nr_args) 531 { 532 __u32 tmp; 533 int err; 534 535 lockdep_assert_held(&ctx->uring_lock); 536 537 if (check_add_overflow(up->offset, nr_args, &tmp)) 538 return -EOVERFLOW; 539 err = io_rsrc_node_switch_start(ctx); 540 if (err) 541 return err; 542 543 switch (type) { 544 case IORING_RSRC_FILE: 545 return __io_sqe_files_update(ctx, up, nr_args); 546 case IORING_RSRC_BUFFER: 547 return __io_sqe_buffers_update(ctx, up, nr_args); 548 } 549 return -EINVAL; 550 } 551 552 int io_register_files_update(struct io_ring_ctx *ctx, void __user *arg, 553 unsigned nr_args) 554 { 555 struct io_uring_rsrc_update2 up; 556 557 if (!nr_args) 558 return -EINVAL; 559 memset(&up, 0, sizeof(up)); 560 if (copy_from_user(&up, arg, sizeof(struct io_uring_rsrc_update))) 561 return -EFAULT; 562 if (up.resv || up.resv2) 563 return -EINVAL; 564 return __io_register_rsrc_update(ctx, IORING_RSRC_FILE, &up, nr_args); 565 } 566 567 int io_register_rsrc_update(struct io_ring_ctx *ctx, void __user *arg, 568 unsigned size, unsigned type) 569 { 570 struct io_uring_rsrc_update2 up; 571 572 if (size != sizeof(up)) 573 return -EINVAL; 574 if (copy_from_user(&up, arg, sizeof(up))) 575 return -EFAULT; 576 if (!up.nr || up.resv || up.resv2) 577 return -EINVAL; 578 return __io_register_rsrc_update(ctx, type, &up, up.nr); 579 } 580 581 __cold int io_register_rsrc(struct io_ring_ctx *ctx, void __user *arg, 582 unsigned int size, unsigned int type) 583 { 584 struct io_uring_rsrc_register rr; 585 586 /* keep it extendible */ 587 if (size != sizeof(rr)) 588 return -EINVAL; 589 590 memset(&rr, 0, sizeof(rr)); 591 if (copy_from_user(&rr, arg, size)) 592 return -EFAULT; 593 if (!rr.nr || rr.resv2) 594 return -EINVAL; 595 if (rr.flags & ~IORING_RSRC_REGISTER_SPARSE) 596 return -EINVAL; 597 598 switch (type) { 599 case IORING_RSRC_FILE: 600 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 601 break; 602 return io_sqe_files_register(ctx, u64_to_user_ptr(rr.data), 603 rr.nr, u64_to_user_ptr(rr.tags)); 604 case IORING_RSRC_BUFFER: 605 if (rr.flags & IORING_RSRC_REGISTER_SPARSE && rr.data) 606 break; 607 return io_sqe_buffers_register(ctx, u64_to_user_ptr(rr.data), 608 rr.nr, u64_to_user_ptr(rr.tags)); 609 } 610 return -EINVAL; 611 } 612 613 int io_files_update_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe) 614 { 615 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 616 617 if (unlikely(req->flags & (REQ_F_FIXED_FILE | REQ_F_BUFFER_SELECT))) 618 return -EINVAL; 619 if (sqe->rw_flags || sqe->splice_fd_in) 620 return -EINVAL; 621 622 up->offset = READ_ONCE(sqe->off); 623 up->nr_args = READ_ONCE(sqe->len); 624 if (!up->nr_args) 625 return -EINVAL; 626 up->arg = READ_ONCE(sqe->addr); 627 return 0; 628 } 629 630 static int io_files_update_with_index_alloc(struct io_kiocb *req, 631 unsigned int issue_flags) 632 { 633 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 634 __s32 __user *fds = u64_to_user_ptr(up->arg); 635 unsigned int done; 636 struct file *file; 637 int ret, fd; 638 639 if (!req->ctx->file_data) 640 return -ENXIO; 641 642 for (done = 0; done < up->nr_args; done++) { 643 if (copy_from_user(&fd, &fds[done], sizeof(fd))) { 644 ret = -EFAULT; 645 break; 646 } 647 648 file = fget(fd); 649 if (!file) { 650 ret = -EBADF; 651 break; 652 } 653 ret = io_fixed_fd_install(req, issue_flags, file, 654 IORING_FILE_INDEX_ALLOC); 655 if (ret < 0) 656 break; 657 if (copy_to_user(&fds[done], &ret, sizeof(ret))) { 658 __io_close_fixed(req->ctx, issue_flags, ret); 659 ret = -EFAULT; 660 break; 661 } 662 } 663 664 if (done) 665 return done; 666 return ret; 667 } 668 669 int io_files_update(struct io_kiocb *req, unsigned int issue_flags) 670 { 671 struct io_rsrc_update *up = io_kiocb_to_cmd(req, struct io_rsrc_update); 672 struct io_ring_ctx *ctx = req->ctx; 673 struct io_uring_rsrc_update2 up2; 674 int ret; 675 676 up2.offset = up->offset; 677 up2.data = up->arg; 678 up2.nr = 0; 679 up2.tags = 0; 680 up2.resv = 0; 681 up2.resv2 = 0; 682 683 if (up->offset == IORING_FILE_INDEX_ALLOC) { 684 ret = io_files_update_with_index_alloc(req, issue_flags); 685 } else { 686 io_ring_submit_lock(ctx, issue_flags); 687 ret = __io_register_rsrc_update(ctx, IORING_RSRC_FILE, 688 &up2, up->nr_args); 689 io_ring_submit_unlock(ctx, issue_flags); 690 } 691 692 if (ret < 0) 693 req_set_fail(req); 694 io_req_set_res(req, ret, 0); 695 return IOU_OK; 696 } 697 698 int io_queue_rsrc_removal(struct io_rsrc_data *data, unsigned idx, 699 struct io_rsrc_node *node, void *rsrc) 700 { 701 u64 *tag_slot = io_get_tag_slot(data, idx); 702 struct io_rsrc_put *prsrc; 703 bool inline_item = true; 704 705 if (!node->inline_items) { 706 prsrc = &node->item; 707 node->inline_items++; 708 } else { 709 prsrc = kzalloc(sizeof(*prsrc), GFP_KERNEL); 710 if (!prsrc) 711 return -ENOMEM; 712 inline_item = false; 713 } 714 715 prsrc->tag = *tag_slot; 716 *tag_slot = 0; 717 prsrc->rsrc = rsrc; 718 if (!inline_item) 719 list_add(&prsrc->list, &node->item_list); 720 return 0; 721 } 722 723 void __io_sqe_files_unregister(struct io_ring_ctx *ctx) 724 { 725 int i; 726 727 for (i = 0; i < ctx->nr_user_files; i++) { 728 struct file *file = io_file_from_index(&ctx->file_table, i); 729 730 /* skip scm accounted files, they'll be freed by ->ring_sock */ 731 if (!file || io_file_need_scm(file)) 732 continue; 733 io_file_bitmap_clear(&ctx->file_table, i); 734 fput(file); 735 } 736 737 #if defined(CONFIG_UNIX) 738 if (ctx->ring_sock) { 739 struct sock *sock = ctx->ring_sock->sk; 740 struct sk_buff *skb; 741 742 while ((skb = skb_dequeue(&sock->sk_receive_queue)) != NULL) 743 kfree_skb(skb); 744 } 745 #endif 746 io_free_file_tables(&ctx->file_table); 747 io_file_table_set_alloc_range(ctx, 0, 0); 748 io_rsrc_data_free(ctx->file_data); 749 ctx->file_data = NULL; 750 ctx->nr_user_files = 0; 751 } 752 753 int io_sqe_files_unregister(struct io_ring_ctx *ctx) 754 { 755 unsigned nr = ctx->nr_user_files; 756 int ret; 757 758 if (!ctx->file_data) 759 return -ENXIO; 760 761 /* 762 * Quiesce may unlock ->uring_lock, and while it's not held 763 * prevent new requests using the table. 764 */ 765 ctx->nr_user_files = 0; 766 ret = io_rsrc_ref_quiesce(ctx->file_data, ctx); 767 ctx->nr_user_files = nr; 768 if (!ret) 769 __io_sqe_files_unregister(ctx); 770 return ret; 771 } 772 773 /* 774 * Ensure the UNIX gc is aware of our file set, so we are certain that 775 * the io_uring can be safely unregistered on process exit, even if we have 776 * loops in the file referencing. We account only files that can hold other 777 * files because otherwise they can't form a loop and so are not interesting 778 * for GC. 779 */ 780 int __io_scm_file_account(struct io_ring_ctx *ctx, struct file *file) 781 { 782 #if defined(CONFIG_UNIX) 783 struct sock *sk = ctx->ring_sock->sk; 784 struct sk_buff_head *head = &sk->sk_receive_queue; 785 struct scm_fp_list *fpl; 786 struct sk_buff *skb; 787 788 if (likely(!io_file_need_scm(file))) 789 return 0; 790 791 /* 792 * See if we can merge this file into an existing skb SCM_RIGHTS 793 * file set. If there's no room, fall back to allocating a new skb 794 * and filling it in. 795 */ 796 spin_lock_irq(&head->lock); 797 skb = skb_peek(head); 798 if (skb && UNIXCB(skb).fp->count < SCM_MAX_FD) 799 __skb_unlink(skb, head); 800 else 801 skb = NULL; 802 spin_unlock_irq(&head->lock); 803 804 if (!skb) { 805 fpl = kzalloc(sizeof(*fpl), GFP_KERNEL); 806 if (!fpl) 807 return -ENOMEM; 808 809 skb = alloc_skb(0, GFP_KERNEL); 810 if (!skb) { 811 kfree(fpl); 812 return -ENOMEM; 813 } 814 815 fpl->user = get_uid(current_user()); 816 fpl->max = SCM_MAX_FD; 817 fpl->count = 0; 818 819 UNIXCB(skb).fp = fpl; 820 skb->sk = sk; 821 skb->scm_io_uring = 1; 822 skb->destructor = unix_destruct_scm; 823 refcount_add(skb->truesize, &sk->sk_wmem_alloc); 824 } 825 826 fpl = UNIXCB(skb).fp; 827 fpl->fp[fpl->count++] = get_file(file); 828 unix_inflight(fpl->user, file); 829 skb_queue_head(head, skb); 830 fput(file); 831 #endif 832 return 0; 833 } 834 835 static __cold void io_rsrc_file_scm_put(struct io_ring_ctx *ctx, struct file *file) 836 { 837 #if defined(CONFIG_UNIX) 838 struct sock *sock = ctx->ring_sock->sk; 839 struct sk_buff_head list, *head = &sock->sk_receive_queue; 840 struct sk_buff *skb; 841 int i; 842 843 __skb_queue_head_init(&list); 844 845 /* 846 * Find the skb that holds this file in its SCM_RIGHTS. When found, 847 * remove this entry and rearrange the file array. 848 */ 849 skb = skb_dequeue(head); 850 while (skb) { 851 struct scm_fp_list *fp; 852 853 fp = UNIXCB(skb).fp; 854 for (i = 0; i < fp->count; i++) { 855 int left; 856 857 if (fp->fp[i] != file) 858 continue; 859 860 unix_notinflight(fp->user, fp->fp[i]); 861 left = fp->count - 1 - i; 862 if (left) { 863 memmove(&fp->fp[i], &fp->fp[i + 1], 864 left * sizeof(struct file *)); 865 } 866 fp->count--; 867 if (!fp->count) { 868 kfree_skb(skb); 869 skb = NULL; 870 } else { 871 __skb_queue_tail(&list, skb); 872 } 873 fput(file); 874 file = NULL; 875 break; 876 } 877 878 if (!file) 879 break; 880 881 __skb_queue_tail(&list, skb); 882 883 skb = skb_dequeue(head); 884 } 885 886 if (skb_peek(&list)) { 887 spin_lock_irq(&head->lock); 888 while ((skb = __skb_dequeue(&list)) != NULL) 889 __skb_queue_tail(head, skb); 890 spin_unlock_irq(&head->lock); 891 } 892 #endif 893 } 894 895 static void io_rsrc_file_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 896 { 897 struct file *file = prsrc->file; 898 899 if (likely(!io_file_need_scm(file))) 900 fput(file); 901 else 902 io_rsrc_file_scm_put(ctx, file); 903 } 904 905 int io_sqe_files_register(struct io_ring_ctx *ctx, void __user *arg, 906 unsigned nr_args, u64 __user *tags) 907 { 908 __s32 __user *fds = (__s32 __user *) arg; 909 struct file *file; 910 int fd, ret; 911 unsigned i; 912 913 if (ctx->file_data) 914 return -EBUSY; 915 if (!nr_args) 916 return -EINVAL; 917 if (nr_args > IORING_MAX_FIXED_FILES) 918 return -EMFILE; 919 if (nr_args > rlimit(RLIMIT_NOFILE)) 920 return -EMFILE; 921 ret = io_rsrc_data_alloc(ctx, io_rsrc_file_put, tags, nr_args, 922 &ctx->file_data); 923 if (ret) 924 return ret; 925 926 if (!io_alloc_file_tables(&ctx->file_table, nr_args)) { 927 io_rsrc_data_free(ctx->file_data); 928 ctx->file_data = NULL; 929 return -ENOMEM; 930 } 931 932 for (i = 0; i < nr_args; i++, ctx->nr_user_files++) { 933 struct io_fixed_file *file_slot; 934 935 if (fds && copy_from_user(&fd, &fds[i], sizeof(fd))) { 936 ret = -EFAULT; 937 goto fail; 938 } 939 /* allow sparse sets */ 940 if (!fds || fd == -1) { 941 ret = -EINVAL; 942 if (unlikely(*io_get_tag_slot(ctx->file_data, i))) 943 goto fail; 944 continue; 945 } 946 947 file = fget(fd); 948 ret = -EBADF; 949 if (unlikely(!file)) 950 goto fail; 951 952 /* 953 * Don't allow io_uring instances to be registered. If UNIX 954 * isn't enabled, then this causes a reference cycle and this 955 * instance can never get freed. If UNIX is enabled we'll 956 * handle it just fine, but there's still no point in allowing 957 * a ring fd as it doesn't support regular read/write anyway. 958 */ 959 if (io_is_uring_fops(file)) { 960 fput(file); 961 goto fail; 962 } 963 ret = io_scm_file_account(ctx, file); 964 if (ret) { 965 fput(file); 966 goto fail; 967 } 968 file_slot = io_fixed_file_slot(&ctx->file_table, i); 969 io_fixed_file_set(file_slot, file); 970 io_file_bitmap_set(&ctx->file_table, i); 971 } 972 973 /* default it to the whole table */ 974 io_file_table_set_alloc_range(ctx, 0, ctx->nr_user_files); 975 return 0; 976 fail: 977 __io_sqe_files_unregister(ctx); 978 return ret; 979 } 980 981 static void io_rsrc_buf_put(struct io_ring_ctx *ctx, struct io_rsrc_put *prsrc) 982 { 983 io_buffer_unmap(ctx, &prsrc->buf); 984 prsrc->buf = NULL; 985 } 986 987 void __io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 988 { 989 unsigned int i; 990 991 for (i = 0; i < ctx->nr_user_bufs; i++) 992 io_buffer_unmap(ctx, &ctx->user_bufs[i]); 993 kfree(ctx->user_bufs); 994 io_rsrc_data_free(ctx->buf_data); 995 ctx->user_bufs = NULL; 996 ctx->buf_data = NULL; 997 ctx->nr_user_bufs = 0; 998 } 999 1000 int io_sqe_buffers_unregister(struct io_ring_ctx *ctx) 1001 { 1002 unsigned nr = ctx->nr_user_bufs; 1003 int ret; 1004 1005 if (!ctx->buf_data) 1006 return -ENXIO; 1007 1008 /* 1009 * Quiesce may unlock ->uring_lock, and while it's not held 1010 * prevent new requests using the table. 1011 */ 1012 ctx->nr_user_bufs = 0; 1013 ret = io_rsrc_ref_quiesce(ctx->buf_data, ctx); 1014 ctx->nr_user_bufs = nr; 1015 if (!ret) 1016 __io_sqe_buffers_unregister(ctx); 1017 return ret; 1018 } 1019 1020 /* 1021 * Not super efficient, but this is just a registration time. And we do cache 1022 * the last compound head, so generally we'll only do a full search if we don't 1023 * match that one. 1024 * 1025 * We check if the given compound head page has already been accounted, to 1026 * avoid double accounting it. This allows us to account the full size of the 1027 * page, not just the constituent pages of a huge page. 1028 */ 1029 static bool headpage_already_acct(struct io_ring_ctx *ctx, struct page **pages, 1030 int nr_pages, struct page *hpage) 1031 { 1032 int i, j; 1033 1034 /* check current page array */ 1035 for (i = 0; i < nr_pages; i++) { 1036 if (!PageCompound(pages[i])) 1037 continue; 1038 if (compound_head(pages[i]) == hpage) 1039 return true; 1040 } 1041 1042 /* check previously registered pages */ 1043 for (i = 0; i < ctx->nr_user_bufs; i++) { 1044 struct io_mapped_ubuf *imu = ctx->user_bufs[i]; 1045 1046 for (j = 0; j < imu->nr_bvecs; j++) { 1047 if (!PageCompound(imu->bvec[j].bv_page)) 1048 continue; 1049 if (compound_head(imu->bvec[j].bv_page) == hpage) 1050 return true; 1051 } 1052 } 1053 1054 return false; 1055 } 1056 1057 static int io_buffer_account_pin(struct io_ring_ctx *ctx, struct page **pages, 1058 int nr_pages, struct io_mapped_ubuf *imu, 1059 struct page **last_hpage) 1060 { 1061 int i, ret; 1062 1063 imu->acct_pages = 0; 1064 for (i = 0; i < nr_pages; i++) { 1065 if (!PageCompound(pages[i])) { 1066 imu->acct_pages++; 1067 } else { 1068 struct page *hpage; 1069 1070 hpage = compound_head(pages[i]); 1071 if (hpage == *last_hpage) 1072 continue; 1073 *last_hpage = hpage; 1074 if (headpage_already_acct(ctx, pages, i, hpage)) 1075 continue; 1076 imu->acct_pages += page_size(hpage) >> PAGE_SHIFT; 1077 } 1078 } 1079 1080 if (!imu->acct_pages) 1081 return 0; 1082 1083 ret = io_account_mem(ctx, imu->acct_pages); 1084 if (ret) 1085 imu->acct_pages = 0; 1086 return ret; 1087 } 1088 1089 struct page **io_pin_pages(unsigned long ubuf, unsigned long len, int *npages) 1090 { 1091 unsigned long start, end, nr_pages; 1092 struct vm_area_struct **vmas = NULL; 1093 struct page **pages = NULL; 1094 int i, pret, ret = -ENOMEM; 1095 1096 end = (ubuf + len + PAGE_SIZE - 1) >> PAGE_SHIFT; 1097 start = ubuf >> PAGE_SHIFT; 1098 nr_pages = end - start; 1099 1100 pages = kvmalloc_array(nr_pages, sizeof(struct page *), GFP_KERNEL); 1101 if (!pages) 1102 goto done; 1103 1104 vmas = kvmalloc_array(nr_pages, sizeof(struct vm_area_struct *), 1105 GFP_KERNEL); 1106 if (!vmas) 1107 goto done; 1108 1109 ret = 0; 1110 mmap_read_lock(current->mm); 1111 pret = pin_user_pages(ubuf, nr_pages, FOLL_WRITE | FOLL_LONGTERM, 1112 pages, vmas); 1113 if (pret == nr_pages) { 1114 struct file *file = vmas[0]->vm_file; 1115 1116 /* don't support file backed memory */ 1117 for (i = 0; i < nr_pages; i++) { 1118 if (vmas[i]->vm_file != file) { 1119 ret = -EINVAL; 1120 break; 1121 } 1122 if (!file) 1123 continue; 1124 if (!vma_is_shmem(vmas[i]) && !is_file_hugepages(file)) { 1125 ret = -EOPNOTSUPP; 1126 break; 1127 } 1128 } 1129 *npages = nr_pages; 1130 } else { 1131 ret = pret < 0 ? pret : -EFAULT; 1132 } 1133 mmap_read_unlock(current->mm); 1134 if (ret) { 1135 /* 1136 * if we did partial map, or found file backed vmas, 1137 * release any pages we did get 1138 */ 1139 if (pret > 0) 1140 unpin_user_pages(pages, pret); 1141 goto done; 1142 } 1143 ret = 0; 1144 done: 1145 kvfree(vmas); 1146 if (ret < 0) { 1147 kvfree(pages); 1148 pages = ERR_PTR(ret); 1149 } 1150 return pages; 1151 } 1152 1153 static int io_sqe_buffer_register(struct io_ring_ctx *ctx, struct iovec *iov, 1154 struct io_mapped_ubuf **pimu, 1155 struct page **last_hpage) 1156 { 1157 struct io_mapped_ubuf *imu = NULL; 1158 struct page **pages = NULL; 1159 unsigned long off; 1160 size_t size; 1161 int ret, nr_pages, i; 1162 struct folio *folio = NULL; 1163 1164 *pimu = ctx->dummy_ubuf; 1165 if (!iov->iov_base) 1166 return 0; 1167 1168 ret = -ENOMEM; 1169 pages = io_pin_pages((unsigned long) iov->iov_base, iov->iov_len, 1170 &nr_pages); 1171 if (IS_ERR(pages)) { 1172 ret = PTR_ERR(pages); 1173 pages = NULL; 1174 goto done; 1175 } 1176 1177 /* If it's a huge page, try to coalesce them into a single bvec entry */ 1178 if (nr_pages > 1) { 1179 folio = page_folio(pages[0]); 1180 for (i = 1; i < nr_pages; i++) { 1181 if (page_folio(pages[i]) != folio) { 1182 folio = NULL; 1183 break; 1184 } 1185 } 1186 if (folio) { 1187 /* 1188 * The pages are bound to the folio, it doesn't 1189 * actually unpin them but drops all but one reference, 1190 * which is usually put down by io_buffer_unmap(). 1191 * Note, needs a better helper. 1192 */ 1193 unpin_user_pages(&pages[1], nr_pages - 1); 1194 nr_pages = 1; 1195 } 1196 } 1197 1198 imu = kvmalloc(struct_size(imu, bvec, nr_pages), GFP_KERNEL); 1199 if (!imu) 1200 goto done; 1201 1202 ret = io_buffer_account_pin(ctx, pages, nr_pages, imu, last_hpage); 1203 if (ret) { 1204 unpin_user_pages(pages, nr_pages); 1205 goto done; 1206 } 1207 1208 off = (unsigned long) iov->iov_base & ~PAGE_MASK; 1209 size = iov->iov_len; 1210 /* store original address for later verification */ 1211 imu->ubuf = (unsigned long) iov->iov_base; 1212 imu->ubuf_end = imu->ubuf + iov->iov_len; 1213 imu->nr_bvecs = nr_pages; 1214 *pimu = imu; 1215 ret = 0; 1216 1217 if (folio) { 1218 bvec_set_page(&imu->bvec[0], pages[0], size, off); 1219 goto done; 1220 } 1221 for (i = 0; i < nr_pages; i++) { 1222 size_t vec_len; 1223 1224 vec_len = min_t(size_t, size, PAGE_SIZE - off); 1225 bvec_set_page(&imu->bvec[i], pages[i], vec_len, off); 1226 off = 0; 1227 size -= vec_len; 1228 } 1229 done: 1230 if (ret) 1231 kvfree(imu); 1232 kvfree(pages); 1233 return ret; 1234 } 1235 1236 static int io_buffers_map_alloc(struct io_ring_ctx *ctx, unsigned int nr_args) 1237 { 1238 ctx->user_bufs = kcalloc(nr_args, sizeof(*ctx->user_bufs), GFP_KERNEL); 1239 return ctx->user_bufs ? 0 : -ENOMEM; 1240 } 1241 1242 int io_sqe_buffers_register(struct io_ring_ctx *ctx, void __user *arg, 1243 unsigned int nr_args, u64 __user *tags) 1244 { 1245 struct page *last_hpage = NULL; 1246 struct io_rsrc_data *data; 1247 int i, ret; 1248 struct iovec iov; 1249 1250 BUILD_BUG_ON(IORING_MAX_REG_BUFFERS >= (1u << 16)); 1251 1252 if (ctx->user_bufs) 1253 return -EBUSY; 1254 if (!nr_args || nr_args > IORING_MAX_REG_BUFFERS) 1255 return -EINVAL; 1256 ret = io_rsrc_data_alloc(ctx, io_rsrc_buf_put, tags, nr_args, &data); 1257 if (ret) 1258 return ret; 1259 ret = io_buffers_map_alloc(ctx, nr_args); 1260 if (ret) { 1261 io_rsrc_data_free(data); 1262 return ret; 1263 } 1264 1265 for (i = 0; i < nr_args; i++, ctx->nr_user_bufs++) { 1266 if (arg) { 1267 ret = io_copy_iov(ctx, &iov, arg, i); 1268 if (ret) 1269 break; 1270 ret = io_buffer_validate(&iov); 1271 if (ret) 1272 break; 1273 } else { 1274 memset(&iov, 0, sizeof(iov)); 1275 } 1276 1277 if (!iov.iov_base && *io_get_tag_slot(data, i)) { 1278 ret = -EINVAL; 1279 break; 1280 } 1281 1282 ret = io_sqe_buffer_register(ctx, &iov, &ctx->user_bufs[i], 1283 &last_hpage); 1284 if (ret) 1285 break; 1286 } 1287 1288 WARN_ON_ONCE(ctx->buf_data); 1289 1290 ctx->buf_data = data; 1291 if (ret) 1292 __io_sqe_buffers_unregister(ctx); 1293 return ret; 1294 } 1295 1296 int io_import_fixed(int ddir, struct iov_iter *iter, 1297 struct io_mapped_ubuf *imu, 1298 u64 buf_addr, size_t len) 1299 { 1300 u64 buf_end; 1301 size_t offset; 1302 1303 if (WARN_ON_ONCE(!imu)) 1304 return -EFAULT; 1305 if (unlikely(check_add_overflow(buf_addr, (u64)len, &buf_end))) 1306 return -EFAULT; 1307 /* not inside the mapped region */ 1308 if (unlikely(buf_addr < imu->ubuf || buf_end > imu->ubuf_end)) 1309 return -EFAULT; 1310 1311 /* 1312 * Might not be a start of buffer, set size appropriately 1313 * and advance us to the beginning. 1314 */ 1315 offset = buf_addr - imu->ubuf; 1316 iov_iter_bvec(iter, ddir, imu->bvec, imu->nr_bvecs, offset + len); 1317 1318 if (offset) { 1319 /* 1320 * Don't use iov_iter_advance() here, as it's really slow for 1321 * using the latter parts of a big fixed buffer - it iterates 1322 * over each segment manually. We can cheat a bit here, because 1323 * we know that: 1324 * 1325 * 1) it's a BVEC iter, we set it up 1326 * 2) all bvecs are PAGE_SIZE in size, except potentially the 1327 * first and last bvec 1328 * 1329 * So just find our index, and adjust the iterator afterwards. 1330 * If the offset is within the first bvec (or the whole first 1331 * bvec, just use iov_iter_advance(). This makes it easier 1332 * since we can just skip the first segment, which may not 1333 * be PAGE_SIZE aligned. 1334 */ 1335 const struct bio_vec *bvec = imu->bvec; 1336 1337 if (offset <= bvec->bv_len) { 1338 /* 1339 * Note, huge pages buffers consists of one large 1340 * bvec entry and should always go this way. The other 1341 * branch doesn't expect non PAGE_SIZE'd chunks. 1342 */ 1343 iter->bvec = bvec; 1344 iter->nr_segs = bvec->bv_len; 1345 iter->count -= offset; 1346 iter->iov_offset = offset; 1347 } else { 1348 unsigned long seg_skip; 1349 1350 /* skip first vec */ 1351 offset -= bvec->bv_len; 1352 seg_skip = 1 + (offset >> PAGE_SHIFT); 1353 1354 iter->bvec = bvec + seg_skip; 1355 iter->nr_segs -= seg_skip; 1356 iter->count -= bvec->bv_len + offset; 1357 iter->iov_offset = offset & ~PAGE_MASK; 1358 } 1359 } 1360 1361 return 0; 1362 } 1363